Table of Contents

List of Figures. 3

List of Tables. 4

Problem 1: Clustering. 5

Executive Summary. 5

1.1 Read the data, do the necessary initial steps, and exploratory data analysis (Univariate, Bi-variate, and multivariate analysis). 5

Sample of the Bank Marketing Dataset. 5

Data Dictionary. 5

Checking the types of variables in the dataset. 5

Descriptive Statistics of the dataset 6

Distribution of the spending variable. 6

Distribution of the advance_payments variable. 7

Distribution of the probability_of_full_payment variable. 7

Distribution of the current_balance variable. 8

Distribution of the credit_limit variable. 8

Distribution of the min_payment_amt variable. 9

Distribution of the max_spent_in_single_shopping variable. 9

Inference. 10

Correlation Heatmap of continuous variables. 10

1.2  Do you think scaling is necessary for clustering in this case? Justify. 10

1.3 Apply hierarchical clustering to scaled data. Identify the number of optimum clusters using Dendrogram and briefly describe them. 11

Visualising Customer Segmentation. 12

1.4 Apply K-Means clustering on scaled data and determine optimum clusters. Apply elbow curve and silhouette score. Explain the results properly. Interpret and write inferences on the finalized clusters. 12

Elbow Method. 12

Silhouette Method. 13

Visualising Customer Segmentation. 14

1.5 Describe cluster profiles for the clusters defined. Recommend different promotional strategies for different clusters. 14

Inferences and Recommendations. 15

Problem 2: CART-RF-ANN.. 15

Executive Summary. 15

2.1 Read the data, do the necessary initial steps, and exploratory data analysis (Univariate, Bi-variate, and multivariate analysis). 15

Sample of the Insurance Dataset. 15

Data Dictionary. 16

Checking the types of variables in the dataset. 16

Descriptive Statistics. 16

Distribution of the Age variable. 17

Distribution of the Agency_Code variable. 18

Distribution of the Type variable. 18

Distribution of the Claimed variable. 19

Distribution of the Commision variable. 19

Distribution of the Channel variable. 20

Distribution of the Duration variable. 20

Distribution of the Sales variable. 21

Distribution of the Product Name variable. 21

Distribution of the Destination variable. 22

Bivariate Analysis of Categorical variables with target variable. 22

Bivariate Analysis of Continuous variables. 23

2.2 Data Split: Split the data into test and train, build classification model CART, Random Forest, Artificial Neural Network. 23

Converting Object variables to Categorical variables. 23

Splitting the data into train and test sets. 23

Scaling the data. 23

CART Model 24

Random Forest Model 24

ANN Model 25

2.3 Performance Metrics: Comment and Check the performance of Predictions on Train and Test sets using Accuracy, Confusion Matrix, Plot ROC curve and get ROC_AUC score, classification reports for each model. 25

Performance Metrics for CART Model. 25

Performance Metrics for Random Forest Model. 27

Performance Metrics for ANN Model. 28

2.4 Final Model: Compare all the models and write an inference which model is best/optimized. 30

2.5 Inference: Based on the whole Analysis, what are the business insights and recommendations. 31

 

 

 

List of Figures

Figure 1: Distribution of the spending variable. 6

Figure 2: Distribution of the advance_payments variable. 7

Figure 3: Distribution of the probability_of_full_payment variable. 7

Figure 4: Distribution of the current_balance variable. 8

Figure 5: Distribution of the credit_limit variable. 8

Figure 6: Distribution of the min_payment_amt variable. 9

Figure 7: Distribution of the max_spent_in_single_shopping variable. 9

Figure 8: Correlation Heatmap. 10

Figure 9: Dendrogram.. 11

Figure 10: Cluster plot for 2 clusters. 12

Figure 11: Elbow Plot for K-Means Clustering. 12

Figure 12: Silhouette Plot. 13

Figure 13: Cluster plot for 3 clusters. 14

Figure 14: Distribution of the Age variable. 16

Figure 15: Distribution of the Agency_Code variable. 17

Figure 16: Distribution of the Type variable. 17

Figure 17: Distribution of the Claimed variable. 18

Figure 18: Distribution of the Commision variable. 18

Figure 19: Distribution of the Channel variable. 19

Figure 20: Distribution of the Duration variable. 19

Figure 21: Distribution of the Sales variable. 20

Figure 22: Distribution of the Product Name variable. 20

Figure 23: Distribution of the Destination variable. 21

Figure 24: Bivariate Analysis of Categorical variables with target variable. 21

Figure 25: Bivariate Analysis of Continuous variables. 22

Figure 26: ROC curves for the CART model 25

Figure 27: ROC Curves for the Random Forest Model 26

Figure 28: ROC Curves for the ANN Model 28

List of Tables

Table 1: Bank Marketing Dataset.................................................................................................................. 5

Table 2: Descriptive Statistics of the dataset................................................................................................ 6

Table 3: Mean and Standard deviation of the data.................................................................................... 10

Table 4: Scaled Bank Marketing Dataset.................................................................................................... 11

Table 5: Sample of Original dataset with Clusters...................................................................................... 13

Table 6: Cluster Profiles................................................................................................................................ 14

Table 7: Sample of the Insurance Dataset................................................................................................... 15

Table 8: Descriptive Statistics of the dataset.............................................................................................. 16

Table 9: Training set after pre-processing................................................................................................... 23

Table 10: Feature importance for the CART model.................................................................................... 23

Table 11: Feature importance for the Random Forest model.................................................................... 23

Table 12: Performance Metrics for Training data...................................................................................... 29

Table 13: Performance Metrics for Testing data........................................................................................ 29

 

A leading bank wants to develop a customer segmentation to give promotional offers to its customers. They collected a sample that summarizes the activities of users during the past few months. You are given the task to identify the segments based on credit card usage.

1.1 Read the data, do the necessary initial steps, and exploratory data analysis (Univariate, Bi-variate, and multivariate analysis).

Sample of the Bank Marketing Dataset.

Table 1 : Bank Marketing Dataset

Table 2 : Descriptive Statistics of the dataset

Looking at the mean values from the above 5-point summary, we see that on an average customers spend approximately 14847 per month. They pay 1455 in advance on average. The average probability that a customer will pay the bank in full is 0.871.

The average current balance for all customers is around 5628, while the credit limit is around 32586. From all the customers, the minimum amount paid for monthly purchases is 370 on average while maximum amount spent in one shopping is 5408.

Table 6 : Cluster Profiles

Inferences and Recommendations

1. For Cluster 0, the mean amount spent by the customers per month is 14438 while the amount paid in advanced is 1433. The average probability of full payment is 88.2%. The balance left in account is 5515 and the credit limit is 32590. The minimum amount paid by the customers is 270 while maximum amount spent in one purchase is 5121. This group represents customers who spend moderately. We can increase credit limit or can lower interest rate by promoting premium cards/loyalty cards to increase spending.
2. For Cluster 1, the mean amount spent by the customers per month is 11857 while the amount paid in advanced is 1325. The average probability of full payment is 84.8%. The balance left in account is 5232 and the credit limit is 28500. The minimum amount paid by the customers is 474 while maximum amount spent in one purchase is 5102. This group represents customers who have low spending power. We can promote cards with offers such as zero annual charges and providing them with benefits such as free coupons or cashback rewards and fee waivers on a variety of places.
3. For Cluster 2, the mean amount spent by the customers per month is 18495 while the amount paid in advanced is 1620. The average probability of full payment is 88.4%. The balance left in account is 6176 and the credit limit is 36980. The minimum amount paid by the customers is 363 while maximum amount spent in one purchase is 6042. This group represents customers who have high spending power. We can offer reward points or discounts on their next big transaction to improve their loyalty.

Problem 2: CART-RF-ANN

Executive Summary

An Insurance firm providing tour insurance is facing higher claim frequency. The management decides to collect data from the past few years. We are assigned the task to make a model which predicts the claim status and provide recommendations to management. Models like CART, RF & ANN are used and the models' performances on train and test sets are compared.

2.1 Read the data, do the necessary initial steps, and exploratory data analysis (Univariate, Bi-variate, and multivariate analysis).

Sample of the Insurance Dataset.

Table 7 : Sample of the Insurance Dataset.

Data Dictionary

1. Claimed: Claim Status (Target)
2. Agency_Code: Code of tour firm
3. Type: Type of tour insurance firms
4. Channel: Distribution channel of tour insurance agencies
5. Product Name: Name of the tour insurance products
6. Duration: Duration of the tour (in days)
7. Destination: Destination of the tour
8. Sales: Amount worth of sales per customer in procuring tour insurance policies in rupees (in 100’s)
9. Commision: The commission received for tour insurance firm (in percentage of sales)
10. Age: Age of insured

Checking the types of variables in the dataset.

<class 'pandas.core.frame.DataFrame'>

RangeIndex: 3000 entries, 0 to 2999

Data columns (total 10 columns):

 #   Column        Non-Null Count  Dtype 

---  ------        --------------  ----- 

 0   Age           3000 non-null   int64 

 1   Agency_Code   3000 non-null   object

 2   Type          3000 non-null   object

 3   Claimed       3000 non-null   object

 4   Commision     3000 non-null   float64

 5   Channel       3000 non-null   object

 6   Duration      3000 non-null   int64 

 7   Sales         3000 non-null   float64

 8   Product Name  3000 non-null   object

 9   Destination   3000 non-null   object

dtypes: float64(2), int64(2), object(6)

memory usage: 234.5+ KB

From the above output we can see that:

·        There are 3000 observations of different individuals in the data.

·        There are 9 independent variables in the dataset.

·        The Claimed column is the target variable.

·        There are 4 numeric columns and 6 categorical (object) columns.

·        The dataset does not have any missing values.

Descriptive Statistics

Table 8 : Descriptive Statistics of the dataset.

In the Insurance dataset, the average age of the customer is 38.09. The commission received for the tour insurance is 14.52% on the sales amount. The average sales amount per customer is Rs. 6024.99. The average duration of the tour for all customers is 70 days. Also, the minimum duration of travel is -1 day, which is not possible. This seems to be bad data which should be taken care of.

There are total 4 tour agencies of which EPX is the most frequent. Out of the 2 types of agencies, Travel Agency occur the most in the data. Also, there are 2 channels, 5 products and 3 destinations provided by the tour insurance firm in the data.

Distribution of the Age variable.

Skewness: 1.1497

Figure 14 : Distribution of the Age variable.

The data in the Age variable has high positive skewness. There are also many outliers present in the variable.

Distribution of the Agency_Code variable.

Figure 15 : Distribution of the Agency_Code variable.

The agency EPX has the highest number of customers while agency JZI has the lowest number of customers in the dataset.

Distribution of the Type variable.

Figure 16 : Distribution of the Type variable.

There are more Travel Agency insurance firms than Airlines travel insurance firms in the dataset.

Distribution of the Claimed variable.

Figure 17 : Distribution of the Claimed variable.

In the given dataset, highest number of customers have not claimed the insurance. Hence, the dataset is imbalanced.

Distribution of the Commision variable.

Skewness: 3.1489

Figure 18 : Distribution of the Commision variable.

The data in the Commision variable has very high positive skewness. There are also many outliers present in the variable.

Distribution of the Channel variable.

Figure 19 : Distribution of the Channel variable.

Almost all of the insurances in the dataset are distributed online. Only few customers have taken the insurance in offline mode.

Distribution of the Duration variable.

Skewness: 13.7847

Figure 20 : Distribution of the Duration variable.

The data in the Duration variable has very high positive skewness. There are also many outliers present in the variable. The high skewness is due to one customer who has a travel duration of more than 4000 days.

 

Distribution of the Sales variable.

Skewness: 2.3811

Figure 21 : Distribution of the Sales variable.

The data in the Sales variable is positively skewed. There are also many outliers present in the variable.

Distribution of the Product Name variable.

Figure 22 : Distribution of the Product Name variable.

From the above figure we see that, highest number of customers have taken Customised insurance plans. The Cancellation plan and the Bronze plan also seem to be popular among customers. The Gold plan is the least popular plan.

Distribution of the Destination variable.

Figure 23 : Distribution of the Destination variable.

The Asia continent seem to be the most popular travel destination among customers in the dataset.

Bivariate Analysis of Categorical variables with target variable.

Figure 24 : Bivariate Analysis of Categorical variables with target variable.

From the above plot we see that, for the agency with code C2B, a greater number of customers have claimed their insurance than those who have not claimed. Also, for the Airlines travel firms, the number of customers who have claimed and who have not claimed are almost the same. Also, most of the customers who have taken Silver plan have claimed their insurance money.

Bivariate Analysis of Continuous variables.

Figure 25 : Bivariate Analysis of Continuous variables.

From the above Correlation plot we see that, the Sales variable is highly correlated with the Commision variable. The Duration variable is moderately correlated to the Sales and Commision variables. Other than that, there is no correlation between other pairs of variables.

2.2 Data Split: Split the data into test and train, build classification model CART, Random Forest, Artificial Neural Network

Converting Object variables to Categorical variables.

The object variables in the dataset were converted to integers using label encoding because most of the machine learning models require that the data be of numeric type.

Splitting the data into train and test sets.

The data was then split into the training and test sets. This is required because we have to make sure that the model generalizes well to unseen data. Therefore, we train the model using the training set and then evaluate it on the test set. For small datasets, the size of the test set is taken to be 20-30% of the dataset. Here we have taken 30% of the data in the test set. Also, the stratify argument is used to make sure that the proportions of labels in the dependent variable is the same in both the training as well as the test set.

Scaling the data.

As the variables in the Insurance dataset have different scales, we also scale the data using StandardScaler function. Scaling is not required for the CART and Random Forest models but it is necessary for Artificial Neural Networks because it is a weight-based model.

Table 9 : Training set after pre-processing

CART Model

Best parameters found for the DecisionTreeClassifier using GridSearchCV after trial and error are:

{'max_depth': 4, 'min_samples_leaf': 9, 'min_samples_split': 25}

Here, the max_depth parameter determines the depth of the tree built by the model, the min_samples_leaf parameter means the minimum number of samples that must be present in the leaf nodes and the min_samples_split parameter means the minimum number of samples that must be present in the decision  nodes for splitting.

Table 10 : Feature importance for the CART model

From the above table we see that, the Agency_Code variable has the highest importance followed by the Sales variable in predicting the claimed status of a customer. All other variables have very low or no importance, meaning they are not very good predictors of the dependent variable.

Random Forest Model

Best parameters found for the RandomForestClassifier using GridSearchCV after trial and error are:

{'max_depth': 7, 'max_features': 3, 'min_samples_leaf': 30, 'min_samples_split': 90, 'n_estimators': 101}

Here, the max_features parameter means the maximum number of features to use for a tree and the n_estimators parameter determines the number of trees to build in the random forest. The remaining parameters are the same as CART model.

Table 11 : Feature importance for the Random Forest model

From the above table we see that, the Agency_Code variable has the highest importance followed by the Product Name, Commision and Sales variables in predicting the claimed status of a customer. All other variables have very low importance, meaning they are not very good predictors of the dependent variable. The Random Forest model gives importance to more variables than the CART model because of bootstrapping.

ANN Model

Best parameters found for the MLPClassifier using GridSearchCV after trial and error are:

{'hidden_layer_sizes': 400, 'max_iter': 300, 'tol': 0.0003}

Here, the hidden_layer_sizes parameter determines the number of neurons in a hidden layer as well as the number of hidden layers in the ANN. The max_iter parameter determines the maximum number of iterations allowed for updating the weights. The tol parameter is the tolerance for the loss function for consecutive iterations.

2.3 Performance Metrics: Comment and Check the performance of Predictions on Train and Test sets using Accuracy, Confusion Matrix, Plot ROC curve and get ROC_AUC score, classification reports for each model.

Performance Metrics for CART Model.

Accuracy Score for Training and Testing

Training Accuracy score for the CART model: 0.797

Testing Accuracy score for the CART model: 0.790

From the above output we see that for the CART model, the training and testing accuracies are very close to each other. Testing score is slightly lower than the training score, therefore the model has very slightly overfitted the data.

Confusion Matrix for Training and Testing

Confusion matrix of the train set for the CART model:

      0    1

0  1260  193

1   234  413

From the above confusion matrix, we see that out of the total 647 positive examples in the training data, 413 were classified correctly while the other 234 examples were classified as negative. Here the False Negatives are greater than the False Positives, hence we can say that the Recall is less than the Precision.

Confusion matrix of the test set for the CART model:

     0    1

0  544   79

1  110  167

From the above confusion matrix, we see that out of the total 277 positive examples in the testing data, 167 were classified correctly while the other 110 examples were classified as negative. Here the False Negatives are greater than the False Positives, hence we can say that the Recall is less than the Precision.

ROC AUC Score and the ROC curves for Training and Testing

ROC AUC score of train data for the CART model: 0.836

ROC AUC score of test data for the CART model: 0.781

Figure 26 : ROC curves for the CART model

From the above output we can see that, the AUC score for the train data is more than the score for the test data; this is also an indicator of slight overfitting. An ideal classifier has the ROC curve closer to top-left corner. The ROC curve for the CART model is far above the diagonal line and near the top-left corner, hence it is a good classifier for the given data.

Classification Reports for Training and Testing

Classification report of the train data for the CART model:

              precision    recall  f1-score   support

 

           0       0.84      0.87      0.86      1453

           1       0.68      0.64      0.66       647

 

    accuracy                           0.80      2100

   macro avg       0.76      0.75      0.76      2100

weighted avg       0.79      0.80      0.79      2100

From the above output we see that, the precision is slightly greater than the recall for the positive class in the training data. Here, it is important that the recall should be large because we need the model to correctly predict as many examples as possible which are actually positive i.e., claimed the insurance. Also, the f1-score value is not very high, hence the overall performance of the model is moderate.

Classification report of the test data for the CART model:

              precision    recall  f1-score   support

 

           0       0.83      0.87      0.85       623

           1       0.68      0.60      0.64       277

 

    accuracy                           0.79       900

   macro avg       0.76      0.74      0.75       900

weighted avg       0.78      0.79      0.79       900

For the test data, the precision is same as that of the training data but the recall has decreased to 0.60. F1-score is very slightly lower than that of the training data. Therefore, we can say that the model generalizes well to unseen data.

Performance Metrics for Random Forest Model.

Accuracy Score for Training and Testing

Training Accuracy score for the Random Forest model: 0.794

Testing Accuracy score for the Random Forest model: 0.782

From the above output we see that for the Random Forest model, the training and testing accuracies are close to each other. Testing score is slightly lower than the training score, therefore the model has slightly overfitted the data.

Confusion Matrix for Training and Testing

Confusion matrix of the train set for the Random Forest model:

      0    1

0  1301  152

1   280  367

From the above confusion matrix, we see that out of the total 647 positive examples in the training data, 367 were classified correctly while the other 280 examples were classified as negative. Here the False Negatives are greater than the False Positives, hence we can say that the Recall is less than the Precision.

Confusion matrix of the test set for the Random Forest model:

     0    1

0  562   61

1  135  142

From the above confusion matrix, we see that out of the total 277 positive examples in the testing data, 142 were classified correctly while the other 135 examples were classified as negative. Here the False Negatives are greater than the False Positives, hence we can say that the Recall is less than the Precision.

ROC AUC Score and the ROC curves for Training and Testing

ROC AUC score of train data for the Random Forest model: 0.846

ROC AUC score of test data for the Random Forest model: 0.802

Figure 27 : ROC Curves for the Random Forest Model

From the above output we can see that, the AUC score for the train data is more than the score for the test data; this is also an indicator of overfitting. The ROC curve for the Random Forest model is far above the diagonal line and near the top-left corner, hence it is a good classifier for the given data.

Classification Reports for Training and Testing

Classification report of the train data for the Random Forest model:

              precision    recall  f1-score   support

 

           0       0.82      0.90      0.86      1453

           1       0.71      0.57      0.63       647

 

    accuracy                           0.79      2100

   macro avg       0.77      0.73      0.74      2100

weighted avg       0.79      0.79      0.79      2100

From the above output we see that, the precision is greater than the recall for the positive class in the training data. Here, it is important that the recall should be large because we need the model to correctly predict as many examples as possible which are actually positive i.e., claimed the insurance. Also, the f1-score value is not very high, hence the overall performance of the model is moderate.

Classification report of the test data for the Random Forest model:

              precision    recall  f1-score   support

 

           0       0.81      0.90      0.85       623

           1       0.70      0.51      0.59       277

 

    accuracy                           0.78       900

   macro avg       0.75      0.71      0.72       900

weighted avg       0.77      0.78      0.77       900

For the test data, the precision is almost same as that of the training data but the recall has decreased to 0.51. F1-score is slightly lower than that of the training data. Therefore, we can say that the model generalizes well to unseen data.

Performance Metrics for ANN Model.

Accuracy Score for Training and Testing

Training Accuracy score for the ANN model: 0.812

Testing Accuracy score for the ANN model: 0.773

From the above output we see that for the ANN model, the training and testing accuracies are different from  each other. Testing score is lower than the training score, therefore the model has overfitted the data.

Confusion Matrix for Training and Testing

Confusion matrix of the train set for the ANN model:

      0    1

0  1291  162

1   233  414

From the above confusion matrix, we see that out of the total 647 positive examples in the training data, 414 were classified correctly while the other 233 examples were classified as negative. Here the False Negatives are greater than the False Positives, hence we can say that the Recall is less than the Precision.

Confusion matrix of the test set for the ANN model:

     0    1

0  555   68

1  136  141

From the above confusion matrix, we see that out of the total 277 positive examples in the testing data, 141 were classified correctly while the other 136 examples were classified as negative. Here the False Negatives are greater than the False Positives, hence we can say that the Recall is less than the Precision.

ROC AUC Score and the ROC curves for Training and Testing

ROC AUC score of train data for the ANN model: 0.869

ROC AUC score of test data for the ANN model: 0.797

Figure 28 : ROC Curves for the ANN Model

From the above output we can see that, the AUC score for the train data is more than the score for the test data; this is also an indicator of overfitting. The ROC curve for the ANN model is far above the diagonal line and near the top-left corner, hence it is a good classifier for the given data. The gap between the training and testing curve larger than the other models.

Classification Reports for Training and Testing

Classification report of the train data for the ANN model:

              precision    recall  f1-score   support

 

           0       0.85      0.89      0.87      1453

           1       0.72      0.64      0.68       647

 

    accuracy                           0.81      2100

   macro avg       0.78      0.76      0.77      2100

weighted avg       0.81      0.81      0.81      2100

From the above output we see that, the precision is greater than the recall for the positive class in the training data. Here, it is important that the recall should be large because we need the model to correctly predict as many examples as possible which are actually positive i.e., claimed the insurance. Also, the f1-score value is not very high, hence the overall performance of the model is moderate.

Classification report of the test data for the ANN model:

              precision    recall  f1-score   support

 

           0       0.80      0.89      0.84       623

           1       0.67      0.51      0.58       277

 

    accuracy                           0.77       900

   macro avg       0.74      0.70      0.71       900

weighted avg       0.76      0.77      0.76       900

For the test data, the precision has reduced to 0.67 and the recall has decreased to 0.51. F1-score is lower than that of the training data. Therefore, we can say that the model does not generalizes well to unseen data. This can be resolved by acquiring more data samples.

2.4 Final Model: Compare all the models and write an inference which model is best/optimized.

Table 12 : Performance Metrics for Training data

Table 13 : Performance Metrics for Testing data

From the above two tables we can see that:

1. The ANN model has the highest accuracy on the training data but lowest accuracy on the test data. The CART and Random Forest model have both the accuracies close to each other. Hence, they have better performance compared to ANN model with respect to accuracy.
2. The Precision for the Random Forest model is most consistent for both training and testing. The ANN model has the highest precision on the training data but lowest precision on the test data.
3. The CART model has the most consistent recall for both the training and testing data. ANN model has the highest recall for training but lowest on testing data.
4. The difference between the f1-score for training and testing for the ANN model is large even though it has the highest score for training data. Again, the CART model has the most consistent f1-score for both the training and testing data.
5. The AUC scores for all the models are high. But the difference between the AUC score for training and testing for the ANN model is large indicating overfitting.

As the recall is the most important metric for this case study and also considering all the other metrics, the CART and the Random Forest model are the best options for model selection. Random Forest model has the highest AUC score on the test data and it will also exhibit less variation on unseen data as it is an ensemble method compared to the CART and ANN model. Hence, Random Forest is the best model for this particular case study.

2.5 Inference: Based on the whole Analysis, what are the business insights and recommendations

For the business problem of a Tour Insurance firm, we have built various models for predicting whether a customer will claim their insurance or not. Models like CART, Random Forest and ANN were evaluated on the training and test datasets and their performances were compared using various performance metrics like Accuracy, Precision, Recall, AUC score, etc.

After comparing the models, it was seen that all of them have performed well on the given data. The ANN model highly overfitted the data. The CART and the Random Forest model performed almost similarly. But as the Random Forest model is an ensemble of many CART models, it’s predictions will be less variable on new data. Therefore, the Random Forest model was determined to be the best model for the business problem.

The Tour insurance firm can make use of the above models as:

1. For Fraud detection – By predicting whether a customer will file for insurance claim or not, the firm can carefully scrutinize the customers who are going to file for claim and detect fraud beforehand.
2. By predicting whether a customer will file for insurance claim or not, the processing of the claims can be made faster and accurate by implementing automation. This will increase customer satisfaction.
3. Referral programs can be implemented for the customers who are going to claim the insurance by providing gift vouchers or providing discounts on their next travel for every successful referral.